Computerized 3D pulmonary architecture analysis

The differential diagnosis of idiopathic interstitial pneumonias (IIPs) with thin-section computed tomography (CT) is an important and particularly difficult task in clinical radiology. The distinction between the two most common entities, idiopathic pulmonary fibrosis (IPF) and nonspecific interstitial pneumonia (NSIP) is especially important because these two entities have substantially different prognoses and usually require different treatment. This project will create an automated computer-aided diagnosis (CAD) system with two main aims: 1) To differentiate between the textural expression of usual interstitial pneumonia (UIP, the histopathological counterpart of IPF) and NSIP in a fully automated fashion while providing increased sensitivity/specificity over today`s differential diagnosis; 2) To allow early diagnosis by performing spatiotemporal analysis of disease progression, yielding novel insights into disease formation. The computational framework we will develop during the project first segments the lung in the CT volume, then computes and selects relevant three-dimensional texture features, and, subsequently differentiates automatically between IPF/UIP and NSIP. The change in the lung`s interstitial architectures is assessed by a classification pipeline based on three-dimensional features that represent the texture sub-types ground-glass opacities, reticular opacities, honeycombing, etc.). The diagnostic power of the system is improved by machine learning approaches, which are used for the automatic selection of the associated texture features. This leads to a classification of IPF/UIP versus NSIP based on an optimal differentiation between the texture sub-types. Tracking the local texture changes between the first examination and all follow-up examinations will yield mathematical models that precisely describe the spatiotemporal progression of the disease, allowing the study of trends within the cohort. The resulting automatic, quantitative analysis of the spatiotemporal progression of IIPs will provide new insights into the disease course characteristics. The proposed CAD framework will enable the quantification and prediction of the progression of IPF/UIP and NSIP, thus aiding the radiologic assessment of IIPs. This method has the potential of being extended to other interstitial lung diseases, considerably impacting clinical applications and future research efforts in this area. The system will also be made available to the research community, providing a common base for lung CT analysis and study comparison.

The PULMARCH project created computational methods that help radiologists to perform diagnosis based on lung imaging data. Diagnosis relies on patterns visible in computed tomography data, but their precise identification is difficult, and the link between their spatial distribution and the underlying disease type and its progression is not yet fully understood. The algorithms developed in PULMARCH allow for the automatic detection of tissue patterns in lung CT data. Novel predictive algorithms compare the progression signature of an individual case with a model population to reliably identify the disease as early as possible. The project integrated research in lung imaging and diagnosis, with advances in machine learning. This made progress in three main directions possible. First, the automatic identification of lung disease patterns from imaging data. Second, the learning with minimal additional supervision. Finally, the determination of the underlying disease based on longitudinal signatures. The impact of these methods is two-fold. Accurate automatic quantification can support individual diagnosis, and large scale clinical trials. The developed machine learning methodology can be the basis for automated image analysis in other anatomical sites and diseases.